ESTRO 2025 - Abstract Book

S3046

Physics - Image acquisition and processing

ESTRO 2025

Results: Observer precision was 2.99 mm, while algorithm precision was 1.56 mm, indicating that algorithms were substantially more precise than observers. However, the accuracy of DIR algorithms averaged 2.88 mm, much higher than their precision, yet similar to observer precision. High inter-observer variability was not attributable to inexperience, as the students' landmarks were, on average, closer to the observer mean than those of the doctors (Table 1). Finally, a weak positive correlation (Pearson R=0.25, Figure 1) was observed between inter-observer and inter-algorithm variability, suggesting that both tend to struggle with similar landmarks, though the relationship is not strong.

Conclusion: For head-and-neck CT landmarking, DIR algorithms were more precise than humans, while algorithm accuracy was comparable to human precision. This suggests that evaluating DIR accuracy based on a single observer's landmarks is insufficient, requiring multiple observers for reliable assessment. Our results however indicate that these observers do not need to be experienced medical personnel.

Keywords: Deformable image registration, uncertainty

References: [1] Seroul, P., & Sarrut, D. (2008). VV: a viewer for the evaluation of 4D image registration. MIDAS Journal, (Medical Image Computing and Computer-Assisted Intervention MICCAI’2008, Workshop - Systems and Architectures for Computer Assisted Interventions), 1–8. http://vv.creatis.insa-lyon.fr